From fb2f8f5b004fe2f768e5daa6fc45d73512badb2a Mon Sep 17 00:00:00 2001
From: Michael Dieringer <65093775+MichaelDieringer@users.noreply.github.com>
Date: Wed, 24 Jun 2026 00:06:42 +0200
Subject: [PATCH] Add CURABIS-MCP-008: AI eval scores must be posted to BC
 posting table

---
 ...-eval-scores-must-be-posted-to-bc-table.md | 98 +++++++++++++++++++
 1 file changed, 98 insertions(+)
 create mode 100644 custom/knowledge/mcp/ai-eval-scores-must-be-posted-to-bc-table.md

diff --git a/custom/knowledge/mcp/ai-eval-scores-must-be-posted-to-bc-table.md b/custom/knowledge/mcp/ai-eval-scores-must-be-posted-to-bc-table.md
new file mode 100644
index 0000000..19ec687
--- /dev/null
+++ b/custom/knowledge/mcp/ai-eval-scores-must-be-posted-to-bc-table.md
@@ -0,0 +1,98 @@
+# CURABIS-MCP-008 — AI eval scores must be posted to the BC posting table
+
+## Rule
+
+When an AI agent completes a hill climbing eval iteration on a BC sub-task, all
+resulting scores — compile result, test score, BCQuality score, F1 score, verdict,
+and model identity — must be posted to the `CUR Project AI Score` table in Business
+Central via the designated MCP tool (`bc_post_ai_score`).
+
+Scores must **not** be stored as:
+- task comments
+- local files or agent memory
+- inline in agent files or knowledge files
+- any other location outside the BC posting table
+
+## Why
+
+The `CUR Project AI Score` table is a **posting table**: one immutable entry per
+iteration, with a clustered key on `Entry No.`. It is the single source of truth for
+hill climbing history on a sub-task.
+
+Storing scores elsewhere breaks this guarantee:
+
+| Alternate location | Problem |
+|---|---|
+| Task comment | 250-char limit, unstructured, not queryable, mixed with human notes |
+| Local file | Session-scoped, repo-specific, invisible to other agents and BC reporting |
+| Agent memory | Volatile, not persisted between sessions |
+| Hard-coded in agent file | Frozen at time of writing, violates CURABIS-MCP-007 pattern |
+
+The BC posting table enables:
+1. Reporting across tasks and projects (MatchRate over time)
+2. The Court reviewing objective score data from Edison
+3. The orchestrator reading prior iterations via `bc_get_ai_scores` to decide verdict
+4. BC users seeing hill climbing progress directly on the sub-task
+
+## Compliant
+
+After each eval iteration, the orchestrator calls:
+
+```
+bc_post_ai_score(
+  projectNo    = "DEV2026-00010",
+  subTaskNo    = "0014",
+  iterationNo  = 3,
+  compile      = true,
+  testScore    = 0.80,
+  bcquality    = 0.86,
+  f1Score      = 0.83,
+  verdict      = "Keep",
+  model        = "claude-sonnet-4-6"
+)
+```
+
+BC sets `Eval DateTime` automatically. The orchestrator may additionally post a
+brief human-readable comment ("Iteration 3: F1=0.83 → Keep") — this is allowed,
+as it communicates progress; the score itself is in BC.
+
+## Non-compliant
+
+```
+# Storing score as task comment only
+bc_add_comment(
+  projectNo = "DEV2026-00010",
+  subTaskNo = "0014",
+  comment   = "Iter 3: compile ✅ tests 4/5 BCQ 6/7 F1=0.83 Keep"
+)
+# → Score is unstructured text. Not queryable. Lost to reporting.
+```
+
+```
+# Storing score in agent file
+## Hill climbing log
+- Iteration 1: F1=0.43 Revert
+- Iteration 2: F1=0.71 Keep
+- Iteration 3: F1=0.83 Keep   ← frozen, session-specific, wrong location
+```
+
+## False positive
+
+An agent that posts a human-readable summary comment **in addition to** calling
+`bc_post_ai_score` is **not** violating this rule. The comment is human
+communication; the score is in BC. Both are permitted.
+
+The violation is using the comment or any other location **instead of** posting to
+the BC table.
+
+## API reference
+
+- Page: `CUR MCP Project AI Scores` (PAG6102906)
+- Entity: `projectAIScores`
+- Publisher: `curabis`, Group: `projectMgmt`, Version: `v2.0`
+- Insert: allowed. Modify: never. Delete: never.
+- `Eval DateTime` is set by BC `OnInsertRecord` — do not pass it.
+
+## Applies to
+
+Agent files that implement hill climbing eval loops on BC sub-tasks.